Release v0.22.3

Highlights

• Added owner-scoped named sessions in Channels to manage up to eight persistent tasks per chat. (#10198)
• Enabled daemon Extension installs to accept absolute local paths while rejecting relative paths and unsupported options. (#10131)
• Updated Web Shell sidebar to display GitHub-style colored state icons for sessions bound to pull requests. (#10258)
• Added an opt-in composer add menu in Web Shell grouping attachments, files, extensions, MCP servers, and skills. (#10379)
• Restored bold text visibility in assistant messages and prevented sidebar session names from being truncated. (#10157)
• Ensured tool summaries are consistently collapsible and separated MCP Apps from other aggregations. (#10231)

Breaking Changes

No known breaking changes.

Complete Change List

Features

• Added a private OpenTelemetry span attribute to report detailed LLM context usage metrics including input categories and remaining capacity. (#10016) by @doudouOUC
• Enabled daemon Extension installs to accept absolute local paths while rejecting relative paths and unsupported options like ref. (#10131) by @callmeYe
• Backfills PR bindings for existing sessions and adds merge-state snapshots to track whether linked pull requests are open, merged, or closed. (#9729) by @wenshao
• Adds an opt-in setting for embedded Web Shell hosts to hide the session source switch, pinning all sidebar catalogs to ordinary task conversations. (#10079) by @dreamWB
• Implements safe projection for reserved [FILE: ...] output markers by redacting malformed or oversized inline markers without uploading files. (#9991) by @qqqys
• Updates Goal prompts to notify the model exactly once when the objective changes, while continuing to assert the current objective on every turn. (#10013) by @qqqys
• Added support for owner-scoped named sessions in Channels, allowing users to manage up to eight persistent named tasks per chat. (#10198) by @doudouOUC
• Updated Web Shell sidebar to display GitHub-style colored state icons for sessions bound to pull requests based on their open, merged, or closed status. (#10258) by @wenshao
• Made deferred review suggestions recoverable by external tooling via a stable invisible marker in the review body artifact. (#9761) by @wenshao
• The /review command now warns when a subsystem's Critical issues repeatedly regrow across rounds by detecting successor chains. (#10010) by @wenshao
• Restored the desktop brand builder skill for the Tauri shell to generate branded packages using brandId and logo inputs. (#10164) by @yiliang114
• Added a standalone daemon session API supporting create, list, resume, archive, and delete operations for top-level sessions. (#10179) by @doudouOUC
• The /review command now remembers the last explicitly used --effort level per project for subsequent runs without arguments. (#9883) by @tlysanhuo
• Long-running review rounds now update their status comment every 10 minutes with elapsed time and last agent activity to distinguish active work from stalled processes. (#9771) by @wenshao
• Local review-fix loops now anchor incremental rounds to working tree content changes instead of commits, reducing token usage by avoiding full tree re-reviews. (#9659) by @wenshao
• Review findings now support an optional fixConstraint field to specify existing facts or limits that suggested fixes must not violate during the repair process. (#10168) by @wenshao
• feat(sdk): Add standalone session APIs (#10294) by @doudouOUC
• feat(cli): OpenTUI migration infra batch — locked deps, dependency-direction arch check, framework-neutral streaming model (#10134) by @chiga0
• Added a configurable Mem0 extension skeleton providing a retrieval-only external-context-mem0 stdio Extension with bounded HTTP request support. (#10149) by @doudouOUC
• Added a new triage gate that automatically identifies and closes duplicate PRs when their changes are already fully resolved by merged fixes. (#10292) by @yiliang114
• Added an opt-in composer add menu to Web Shell that groups attachments, files, extensions, MCP servers, and skills for easier pointer-oriented selection. (#10379) by @ytahdn
• Added an opt-in interactive browser terminal to Web Shell that manages independent PTY sessions per tab with persistent scrollback and state handling. (#9984) by @ytahdn
• Adds the qwen review ab-drive subcommand to run execution-grade verification scripts against PR and base trees with shared or fresh instances. (#9740) by @wenshao
• Confirmed Critical issues now include direction and baseline axes to distinguish between false certifications, failures, regressions, and new surfaces. (#10301) by @wenshao
• Makes Aone a first-class platform for session PR bindings, enabling automatic branch-to-MR mapping and state refresh via the a1 CLI. (#10287) by @wenshao

Bug Fixes

• Fixed Web Shell to correctly restore saved session diffs with accurate line counts even when file bodies were previously truncated. (#10093) by @ytahdn
• Ensured DingTalk rich-text messages preserve all images as ordered attachments across sessions and handle upload failures gracefully. (#9922) by @BenGuanRan
• fix: repair the Windows test lane's standing failures (#10064) by @wenshao
• fix(core): report broadcast delivery failures in send_message (#10081) by @yiliang114
• Moved the autofix scan lane to the persistent self-hosted runner pool to prevent GitHub-hosted runner backlogs from blocking routing. (#10055) by @wenshao
• Restores bold text visibility in assistant messages and prevents session names in the sidebar from being truncated by attention pills. (#10157) by @wenshao
• Prevents large transcript turns from collapsing automatically after pagination completes, ensuring content remains visible until manually closed. (#10086) by @ytahdn
• Makes the delta-resume reconnect test deterministic by waiting for the resumed chunk instead of using a fixed time delay. (#10203) by @qwen-code-dev-bot
• Verifies the previous release tag exists in git before anchoring release notes, falling back to the latest stable tag if missing. (#10167) by @yiliang114
• Exempts off-Linux E2E shards from fatal unhandled vitest RPC-timeout errors while maintaining strict error handling on Linux. (#10112) by @qwen-code-dev-bot
• Ensures tool summaries are consistently collapsible, prevents child tools from opening automatically, and separates MCP Apps from other aggregations. (#10231) by @ytahdn
• Ensures tool results in headless and ACP hosts are stamped with Goal turn permits so they can serve as citable evidence for Goal reasoning. (#10175) by @qqqys
• Removes a duplicate getToolRegistry mock property in telemetry-swap tests to fix TypeScript build errors caused by duplicate object keys. (#10252) by @wenshao
• Fixes TypeScript build failure by removing a duplicate getToolRegistry property from the telemetry swap test configuration mock. (#10243) by @AaronZ345
• Adds qwen3.8-flash and qwen3.8-plus to modality patterns to enable native image and video handling instead of routing through the vision bridge. (#10195) by @yiliang114
• Bounds the goal-runtime startup wait to 5 seconds to prevent hanging and skips unnecessary memory relaunch steps when running under Bun. (#10128) by @chiga0
• Implemented robust process tree reaping for ACP child processes on POSIX and Windows to prevent orphaned processes during shutdown or errors. (#10142) by @doudouOUC
• Ensured model-specific reasoning controls remain available and correctly synchronized for qwen3.8-max before, during, and after session execution. (#10244) by @callmeYe
• Fixed command hook cleanup to reliably reclaim entire process trees on POSIX and Windows when hooks time out or are cancelled. (#10100) by @doudouOUC
• Auto-detection now correctly identifies deepseek-v4-flash-vision-exp, glm-4.6v, glm-5v-turbo, and glm-5.3-flash as vision-capable models. (#10278) by @yiliang114
• Increased autofix agent time budgets to 60 minutes for repair passes and 90 minutes for the runner default to prevent premature task termination. (#10217) by @wenshao
• Updated stream idle timeout error messages to guide users to both environment variables and explicit settings for resolution. (#9896) by @yu-xin-c
• The Web Shell new-session welcome header now features improved typography, spacing, a fade-in animation, and a brand-aligned gradient matching the Qwen logo. (#10230) by @wenshao
• Artifact updatedAt timestamps and sizeBytes now update immediately when content changes, ensuring Web Shell previews and caches reflect the latest file state. (#9929) by @zjgzx1988
• Extension Skills now separate canonical identity from display names, ensuring consistent matching while showing friendly localized labels in the CLI and Web Shell. (#10296) by @callmeYe
• Agent View now preserves queued follow-up messages in shared context across tab switches, preventing data loss when teammates switch views while an agent is busy. (#10102) by @yiliang114
• Stale team reclamation is now generation-safe, aborting deletion if the team configuration has changed since the staleness check to prevent accidental data loss. (#10236) by @yiliang114
• fix(core): report teammate result settled before event bridge attach (#10245) by @yiliang114
• fix(cli): restore goal-less startup broken by the goal-runtime wait refactor (#10290) by @yiliang114
• fix(test): mount channel plugin fixture in Docker (#10318) by @yiliang114
• fix(ci): stop shared-pool hosts exiting all-green E2E runs red (#10325) (#10329) by @qwen-code-dev-bot
• Integration tests now run with an isolated global qwen dir to prevent host configuration and saved memories from affecting test results. (#10337) by @wenshao
• E2E workflows now include a bounded retry for transient sandbox image build failures in the Linux docker leg to improve CI stability. (#10359) by @qwen-code-dev-bot
• Moved the archive action to the session dropdown and disabled it for running sessions to prevent UI overlap and accidental archiving. (#10302) by @wenshao
• Skill toggle endpoints now persist requested names directly to workspace settings without validating against the runtime Skill catalog. (#10077) by @callmeYe
• fix(test): isolate tool-control E2E from shared state (#10340) by @yiliang114
• Wind-down hand-off markers are now stamped only when the turn is successfully delivered to the model to ensure correct budget handling. (#10260) by @qqqys
• Direct DWS final responses now use standard direct-message delivery and recovery paths to ensure messages are not silently dropped. (#10274) by @qqqys
• Fixed an intermittent Java 11 test failure by ensuring HTTP connections close properly after cancel and detach responses in the concurrency test fixture. (#10365) by @yiliang114
• Updated the ACP cron test fixture to route responses based on prompt content instead of request order, preventing failures caused by parallel test interference. (#10374) by @yiliang114
• Configured Linux E2E shards to use isolated disk-backed temporary directories under /var/tmp to prevent /tmp exhaustion errors during test execution. (#10376) by @yiliang114
• Updated the triage workflow to silently exit when a PR is closed during a re-run, preventing unnecessary warnings about missing bot reviews. (#10334) by @yiliang114
• Updated the triage workflow to correctly fail the check step when the response indicates a model-layer API error instead of treating it as success. (#10333) by @yiliang114
• The ack-review-request job now reposts acknowledgements directly under the triggering command and adds reactions to improve visibility. (#10381) by @qqqys
• Hardens Computer Use sessions with real AbortSignal timeouts, typed errors, authorization reconnection, and consistent foreground delivery. (#10319) by @LaZzyMan
• Decouples permissions.allow from tool registration by introducing settings.tools.eager to control schema visibility without affecting auto-approval. (#10098) by @yiliang114

Documentation

• Removes redundant and non-operational text from agent context files to reduce token usage without altering project rules or commands. (#10137) by @DragonnZhang
• docs(external-context): Design a configurable Mem0 provider extension (#10113) by @doudouOUC

Internal Changes

• Fixed telemetry-swap tests by adding the missing getToolRegistry method to the Config mock to prevent initialization errors. (#10220) by @wenshao
• Removed stale references to the deleted packages/desktop directory from CI tooling and documentation following the Electron removal. (#10021) by @yiliang114
• Renamed internal identifiers containing legacy Gemini prefixes to generic names while maintaining deprecated aliases for backward compatibility. (#9900) by @yiliang114
• Centralized worktree configuration derivation to ensure consistent handling of workspace paths and context across all launch paths. (#9915) by @yiliang114
• refactor(core,cli): rename generic Gemini identifiers to Llm (#10124) by @yiliang114
• Fixes the release classifier test stub to work correctly when Node interprets extensionless executables as either CommonJS or ESM. (#10402) by @ytahdn

Full Changelog: v0.22.2...v0.22.3